- Input Layer:
It initializes the model to accept input images of size (image_size, image_size, 3), which corresponds to image height, image width, and 3 color channels (RGB).
- First Convolutional Block:
Conv2D with 16 filters: Applies a 5x5 convolution kernel to extract features such as edges and textures. The use of 16 filters means it will output 16 different feature maps.
BatchNormalization: Normalizes the activations from the previous layer, which helps in accelerating the training process and stabilizing the learning by normalizing the input layer by re-centering and re-scaling.
MaxPooling2D: Reduces the spatial dimensions (height and width) of the input volume to the next layer by taking the maximum value over a 2x2 pooling window. This helps in reducing the computational cost and overfitting by providing an abstracted form of the representation.
Dropout (0.2): Randomly sets the outgoing edges of 20% of the neurons to zero during training, to prevent overfitting.
- Subsequent Convolutional Blocks:
These blocks increase in the number of filters (32, 64, 128, 256). Increasing the number of filters allows the network to capture more complex patterns like textures and shapes.
Each block follows a similar structure: a convolution layer, batch normalization, max pooling, and dropout. This repeated structure helps the network in learning hierarchically more complex features at each level.
Kernel sizes are generally smaller (3x3) in subsequent layers, which is common as deeper layers capture higher-level abstract features where finer granularity is less important.
- Flattening:
The output of the final convolutional layer is flattened (converted from a matrix to a vector), so it can be fed into the dense layers.
- Dense Layers:
Dense Layer with 512 neurons: This layer is fully connected and uses ReLU activation. It serves as a classifier on the features formed by the convolutions and pooling layers.
Dropout (0.2): Again used here to reduce overfitting.
- Output Layer:
Dense layer with 4 neurons: This implies the model is intended for a classification task with 4 classes. The softmax activation function is used to output a probability distribution over the 4 classes.
- Compilation:
The model uses the Adam optimizer, a popular choice for deep learning tasks as it combines the best properties of the AdaGrad and RMSProp algorithms to optimize its weights.
The loss function is categorical_crossentropy, suitable for multi-class classification problems.
The metric used to evaluate the model is accuracy.
